We are using a Lambda workstation with two TITAN RTX gpus. I recently did a fresh install of Ubuntu 20.04 and installed lambda stack which installed the cuda version 11.1. Unfortunately, this version doesn’t work with the latest PyTorch version 1.7.1 at least on the paper… It seems like PyTorch recognizes the the cuda (
torch.cuda.is_available() = True) but it doesn’t work as expected. For example, FP16 models are slower than FP32 which is not the case on my laptop with CUDA 11.0.
Should I somehow downgrade the cuda to 11.0?